JP5847290B2 - Document search apparatus and document search method - Google Patents

Description

  The present invention relates to a document search apparatus and a document search method for searching for fine units such as chapters, sections, and sections of an electronic document.

  For many devices such as home appliances and in-vehicle devices, a paper instruction manual that describes operation methods and troubleshooting methods is attached. Among them, in an information device having a screen, an instruction manual is digitized and can be directly searched and browsed. This makes it possible to browse without having to carry a paper document. On the other hand, an electronic document has a low listability, and it is difficult to search for a content that a user wants to confirm, and it is essential to provide a search function.

  Among the typical conventional search functions, the simplest method is a GREP search method in which a search is performed using a keyword and hit portions are displayed in the order of appearance from the top of the document. Further, there is a logical search method in which a search index is created from a document and a keyword extracted in advance, and a search is performed using a logical expression using the search index to display candidates. Further, since the logical search method cannot define a score representing the degree of association between an input keyword and a search index, there is a best matching search method in which a score is determined by simply inputting a keyword and counting its appearance frequency. Further, a search index with a statistical weight such as tf.idf (term frequency and inverse document frequency) is created from the keyword, and a search is performed by a vector distance (inner product) with the input keyword to display candidates. There is. By providing these search methods, it is possible to search for an electronic document, and it is possible to browse a portion requested by the user to some extent.

The logical search method searches only those that exactly match the search conditions, so there is an advantage that using complex search conditions makes it easier to find a search that matches the user's search intention. There is a demerit that it is easy to lead to search omission if it is not appropriate. In addition, there is a demerit that building a complicated search expression is a high hurdle for general users. Therefore, the most common logical type search is a method in which a plurality of keywords are input and a search result is obtained and presented by an OR logical operation.
On the other hand, the best-matching search method and the statistical search method have the advantage of being able to search without the need for a logical structure in the keyword, while the number of occurrences of the keyword in the document is simply scored or There is a demerit that it is difficult for the user to control because the score is calculated by a value weighted according to the tendency.

  Based on the merits and demerits of these methods, a method that integrates and processes a plurality of search engines has been proposed as a method of taking advantage of both advantages. For example, Patent Document 1 discloses a method of performing a search by separately executing a logical search method and a statistical search method, or a best matching search method and a statistical search method, and logically integrating the results. .

Specifically, only search result candidate information is obtained from a logical search method search engine, and search result candidates and their scores are obtained as information from best matching search method and statistical search method search engines.
When the logical type search method and the statistical type search method are combined, for example, only the result of the logical type search result and the statistical type search result with the document ID overlapping is set as the final result candidate, or the logical type search result and All the document IDs of the statistical search results are used as final result candidates, and the scores of the statistical search results are used for ranking the final results.

  Furthermore, when the best matching search method and the statistical search method are combined, the final results are ranked using the average of scores.

  In addition, in the conventional search method, a table of synonyms and synonyms is created and the keywords in the search conditions are expanded into synonyms and synonyms in order to reduce the case where the search cannot be performed due to differences in the surface layer of keywords. A search method has been proposed.

Japanese Patent Laid-Open No. 10-143530

  Since the conventional document search apparatus and document search method are configured as described above, it is easier to obtain a search result desired by the user than when searching by a single search method. However, in these search methods, the keyword extraction target for creating the search index is the search target document itself, so even if a single search method is used or a combination of multiple search methods is used, the document It is fundamental to search for keywords that appear inside.

  Further, in the actual search scene, the search side must input the search condition in a state where the keyword used in the document is unknown, so that a desired document cannot be drawn. In order to eliminate them, a search by synonym and synonym expansion is performed, so that some improvement can be expected. However, manuals and other manuals often contain technical terms and explanations using special terms for unique functions to ensure accuracy, so it is recommended for general users and beginner users who want to know how to use them. In many cases, it is difficult to know what is searched for as a keyword to obtain a desired explanation. Specifically, terms such as “north reference” or “own vehicle reference” are terms that cannot be imagined by car navigation beginners as terms that indicate the direction of a map for car navigation. When a search is performed under a condition such as “I want to do”, an appropriate keyword does not exist and a desired search result cannot be obtained.

  The present invention has been made to solve the above-described problems, and it is an object of the present invention to present a more appropriate search result for a user's natural language input than a search result obtained by a simple search method. And

  A document search device according to the present invention receives a search index created from a document prepared in advance and an input from a user, and uses the search index to search an item related to the user input from the document. And an utterance estimation model that learns the correspondence between an assumed question that asks the contents of the document and an item in the document that is the answer to the assumed question, and an item that corresponds to a user input answer from the document based on the utterance estimation model Utterance content estimation unit, and a document search result obtained from the document search unit and a document estimation result obtained from the utterance content estimation unit are integrated to generate a final search result. is there.

In the document search method according to the present invention, a user input step in which an input analysis unit receives an input from a user, and a user input from within the document using a search index created from a document prepared in advance by the document search unit. The document search step for searching for items related to the utterance, and the utterance content estimation unit based on the utterance estimation model in which the correspondence between the assumed question that asks the document content and the item in the document that is the answer to the assumed question is learned. The utterance content estimation step for estimating an item corresponding to the user input answer from the document, and the result integration unit obtains the document search result obtained from the document search step and the document estimation result obtained from the utterance content estimation step. Result in a final search result It is intended and a consolidation step.

  According to the present invention, using an utterance estimation model in which a correspondence between a question assuming a user's way of listening and a document item as an answer is learned, it corresponds to a user input answer from within a document. Since the estimated items are integrated with the index search results, it is possible to present more appropriate search results than the results of the simple search method to the user's natural language input. it can.

It is a block diagram which shows the structure of the document search device concerning Embodiment 1 of this invention. 3 is a diagram illustrating an example of a document of the document search device according to Embodiment 1. FIG. It is a figure which shows the example of the keyword list for the document analysis result and search index of the document search device concerning Embodiment 1. It is a figure which shows the example of the collection speech data of the document search apparatus concerning Embodiment 1. FIG. It is a figure which shows the example of the keyword list for the collection speech analysis results and speech estimation model of the document search device concerning Embodiment 1. FIG. 4 is a flowchart showing an operation of creating a search index from a document of the document search device according to the first embodiment. 4 is a flowchart illustrating an operation of creating an utterance estimation model from collected utterance data of the document search device according to the first embodiment. 6 is a flowchart illustrating an operation of creating a final search result from a user input of the document search device according to the first embodiment. It is a figure which shows the example of transition of a user input in the document search device concerning Embodiment 1. FIG. FIG. 10 is a diagram showing a continuation of the user input transition example of FIG. It is a block diagram which shows the structure of the document search device based on Embodiment 2 of this invention. FIG. 10 is a diagram illustrating a document hierarchy of the document search device according to the second embodiment. 10 is a flowchart illustrating an operation of creating a final search result from a user input of the document search device according to the second embodiment. It is a figure which shows the example of transition of a user input in the document search device concerning Embodiment 2. FIG. It is a figure which shows the example of the document of the document search device concerning Embodiment 3 of this invention. It is a figure which shows the example of the document analysis result of the document search device concerning Embodiment 3, and the keyword list for search indexes. It is a figure which shows the example of the collection speech data of the document search device concerning Embodiment 3. FIG. It is a figure which shows the example of the keyword list for the collection speech analysis results and speech estimation model of the document search device concerning Embodiment 3. FIG. FIG. 10 is a diagram illustrating a transition example of user input in the document search device according to the third embodiment. It is a figure which shows the continuation of the example of a transition of the user input of FIG. It is a figure which shows the example of the document of the document search device concerning Embodiment 4 of this invention. It is a figure which shows the example of the document analysis result of the document search device concerning Embodiment 4, and the keyword list for search indexes. It is a figure which shows the example of the collection speech data of the document search device concerning Embodiment 4. FIG. It is a figure which shows the example of the keyword list for the collection speech analysis results and speech estimation model of the document search device concerning Embodiment 4. FIG. FIG. 10 is a diagram illustrating a transition example of user input in the document search device according to the fourth embodiment. It is a figure which shows the continuation of the example of a transition of the user input of FIG.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
Embodiments of the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram showing the configuration of the document search apparatus according to the first embodiment.
Document 1 is text data obtained by digitizing an instruction manual of a product. This document 1 is assumed to be hierarchized to some extent into items such as chapters, sections, and sections according to product functions. The input analysis unit 2 divides the text of the document 1 into morpheme units by a known technique such as morpheme analysis. The document analysis result 3 is data obtained by dividing the document 1 into morphemes by the input analysis unit 2.

The search index creation unit 4 creates a search index 5 from the document analysis result 3. The search index 5 returns items such as specific chapters, sections, and items in the document 1 as search results in response to the keyword input from the document search unit 12.
The collected utterance data 6 is utterance data collected in advance by using a method such as a user questionnaire when the document 1 is used. The method of creating the collected utterance data 6 assumes that a question is generated from a product function written in the document 1 in advance and collected in advance in the form of a questionnaire or the like.
The collected utterance analysis result 7 is data obtained by dividing the collected utterance data 6 into morphemes by the input analysis unit 2.

  The utterance estimation model creation unit 8 performs statistical learning using the morpheme unit of the collected utterance analysis result 7 as a learning unit (feature), and creates an utterance estimation model 9. This utterance estimation model 9 is learning result data for receiving a morpheme string of the collected utterance analysis result 7 as an input and returning an item corresponding to an answer to the question with a score as an utterance content estimation result.

  The user input 10 is data representing an input from the user to the document search apparatus. Here, description will be made assuming that the user input 10 is a text input. The user input analysis result 11 is data obtained by dividing the user input 10 into morphemes by the input analysis unit 2.

The document search unit 12 performs a search using the search index 5 with the user input analysis result 11 as an input, and creates a document search result 13.
The utterance content estimation unit 14 receives the user input analysis result 11 as an input, estimates an item corresponding to this input using the utterance estimation model 9, and acquires the document ID of the item. The document estimation result 15 is data including the document ID estimated by the utterance content estimation unit 14 and its score (described later).

  The result integration unit 16 integrates the document search result 13 and the document estimation result 15 into one search result and outputs it as a final search result 17.

  FIG. 2 is an example of the document 1. The document 1 has a hierarchical structure such as a chapter, a section, and an item, and has a document ID indicating a search result position for each hierarchy. In the example of FIG. 2, the document 1-1 with the document ID “Id — 10_1” also includes text included in the lower data structure. For example, the document 1-2 of “Id — 10_1 — 1” is included in the document 1-1 of “Id — 10_1”.

FIG. 3 shows an example of the keyword analysis results for the document analysis result 3 and the search index 5. “Id — 10 — 1 — 1” is an example of the document analysis result 3-1, and shows the result of performing input analysis by morphological analysis on the document 1-2 of “Id — 10 — 1 — 1” in FIG. In the document analysis result 3-1, the morphological analysis result is delimited by “/”.
The search index data 3-2 indicates an example of data used by the search index creation unit 4 based on the document analysis result 3-1 of “Id — 10 — 1 — 1”. Here, a document ID and a list of general forms (keywords) of independent word morphemes are extracted.

FIG. 4 is an example of the collected utterance data 6. The collected utterance data 6-1 is an example of a question corresponding to a document “Id_10”, the collected utterance data 6-2 is an example of a question corresponding to a document of “Id_10_1”, and the collected utterance data 6-3 is a document of “Id_10_1_1”. It is an example of a question corresponding to. The collected utterance data 6-4 is a question intended to know a specific method of changing the map type. However, since it is a map type that cannot be realized with the product assumed here, it is the same as "Id_10_1_1". It is an example of the collection utterance data which cannot select document ID of a hierarchy.
Note that these collected utterance data 6-1 to 6-4 are examples of question sentences that are assumed to be heard by the user in order to confirm the function of the product.

FIG. 5 is an example of a keyword list for the collected utterance analysis result 7 and the utterance estimation model 9. “Id — 10 — 1 — 1” is an example of the collected utterance analysis result 7-1 and shows the result of performing input analysis by morphological analysis on the text of the collected utterance data 6-1 of “Id — 10 — 1 — 1” in FIG.
The utterance estimation model data 7-2 shows an example of data used by the utterance estimation model creation unit 8 based on the collected utterance analysis result 7-1 of “Id — 10_1 — 1”. Here, a document ID and a list of general forms (keywords) of independent word morphemes are extracted.

Next, the operation of the document search apparatus will be described.
The operation is roughly divided into two processes. One is a creation process for creating the search index 5 and the utterance estimation model 9 from the document 1 and the collected utterance data 6, respectively, and the other is a search process for creating a final search result 17 in response to the user input 10. is there. First, the creation process will be described.

First, a method for creating the search index 5 in the creation process will be described. Here, weighting by tf · idf disclosed in the prior art is assumed.
FIG. 6 is a flowchart showing an operation from creation of the search index 5 from the document 1. As shown in FIG. 2, it is assumed that the document 1 is a pair in which the document ID and the text are associated with each other. For example, in document 1-2, a document ID “Id — 10 — 1 — 1” is associated with the text “vehicle reference. A map with the direction of travel of the vehicle up is displayed.” In step ST1, the input analysis unit 2 sequentially reads the document 1 having this structure, and divides it into morpheme strings by morphological analysis which is a known technique. The result of the morphological analysis of the document 1-2 is the document analysis result 3-1 in FIG. This document analysis result 3-1 shows only the morpheme delimiter “/”, but it actually includes part-of-speech information, a prototype of a utilization word, reading, and the like.

  When the document analysis result 3 is generated for all document IDs, the search index creation unit 4 extracts morphemes (keywords) necessary for creating the search index 5 from all the document analysis results 3 in the subsequent step ST2. Then, a pair of (document ID, keyword list) is created, and a search index 5 weighted by tf · idf based on all pairs is created. A pair of (document ID, keyword list) extracted from the document analysis result 3-1 in FIG. 3 is represented by the same search index data 3-2 in FIG.

  A specific search index creation procedure will not be described, but will be briefly described. First, tf · idf uses the number of keywords included in all document IDs as a vector dimension, assigns each keyword to a vector element, and represents the vector value as a frequency (tf portion). This vector value is weighted so as to meet the heuristic that “keywords that appear in many documents (general words) have low importance, and keywords that appear only in specific documents have high importance” (idf portion). This weighted table is the search index 5.

Next, a process for creating the utterance estimation model 9 will be described.
FIG. 7 is a flowchart showing an operation until the utterance estimation model 9 is created from the collected utterance data 6. The collected utterance data 6 is data in which utterances collected in advance from the user are assigned to the document ID as a response, as shown in the collected utterance data 6-1 to 6-4 in FIG. The method of creating the collected utterance data 6 is data in which sentences describing what is said when a function for each document ID is presented in a questionnaire or the like and the function is to be searched is collected. For example, in the case of presenting the specific content of “Id_10_1_1” of “Id_10_1_1” in FIG. 4 that “the map is displayed with the traveling direction of the own vehicle up”, the utterance as the collected utterance data 6-3 On the other hand, when a high-level concept such as “Id_10” is presented, data such as collected utterance data 6-1 is started, and collected utterance data 6-2 to 6-4 is started. It can be expected that utterances can also be collected. The collected utterance data 6-4 is utterance data having contents outside the functions of the product of the document 1, and in this case, the collected utterance data 6-4 is assigned to an intermediate document ID “Id — 10_1”. The above operation is performed manually in advance, and data having the structure shown in FIG. 4 is prepared.

  In step ST3, the input analysis unit 2 performs morphological analysis of the collected utterance data 6 in the same manner as when the document 1 is input in step ST1. For example, the result of morphological analysis of the collected utterance data 6-3 in FIG. 4 is the collected utterance analysis result 7-1 in FIG. In the subsequent step ST4, the utterance estimation model creating unit 8 extracts a document ID and keyword list as the utterance estimation model data 7-2 and performs a process for creating the utterance estimation model 9 as in step ST2. It is assumed here that the speech estimation model 9 is learned by the maximum entropy method (hereinafter, ME method).

The ME method will not be described in detail, but will be described briefly. The ME method is a method of estimating a document ID when a pair of (document ID, keyword list) is used as learning data and a keyword list is input as an input. The weight of the pair of (document ID, keyword list) is calculated so that it is most likely to occur in the data learned when estimating the document ID from the keyword list (the number of correct answers increases), and the saved data is the utterance This is an estimation model 9.
Keywords are extracted from all the collected utterance analysis results 7 and learned by the ME method to create an utterance estimation model 9. Specifically, the same utterance estimation model data 7-2 in FIG. 5 is extracted from the collected utterance analysis result 7-1 in FIG. 5, and the above learning is performed based on the utterance estimation model data 7-2. Done.

Next, the search process will be described.
FIG. 8 is a flowchart showing the operation from the user input 10 until the final search result 17 is created. 9 and 10 show transition examples in the search process of the user input 10-1 which is an example of the user input 10. FIG. Here, it is assumed that the user input 10 is a text input and the user input 10-1 in FIG. 9 is input. In step ST11, the input analysis unit 2 first receives the user input 10-1, generates a user input analysis result 11-1 through morphological analysis, extracts an independent word from the user input analysis result 11-1, and extracts the keyword list 11 -2 is created. In subsequent step ST12, the utterance content estimation unit 14 uses the keyword list 11-2 as an input to obtain the document estimation result 15-1 of FIG. 10 from the utterance estimation model 9. As shown in FIG. 10, the document estimation results 15-1 are arranged in the order of scores. This score is a value calculated from the pair weight of (document ID, keyword list) stored in the utterance estimation model 9 and has a high degree of association with the user input 10, that is, the user input 10 A high score is given to the document ID suitable as an answer to the question.

When the document estimation result 15-1 is obtained, the document search unit 12 obtains the document search result 13-1 shown in FIG. 10 from the search index 5 by using the keyword list 11-2 as an input at the next step ST 13. As shown in FIG. 10, document search results 13-1 are also arranged in the order of score. This score is a value calculated from the weight of tf · idf stored in the search index 5, and a high score is given to a document ID having a high degree of association with the user input 10.
Note that a known technique may be used for the calculation method of the score of the document estimation result 15 and the score of the document search result 13, and thus description thereof is omitted here.

  When the process of step ST13 ends, the process proceeds to the process of step ST14, and the result integration unit 16 determines whether the maximum score of the document estimation result 15-1 is equal to or greater than a threshold value X (for example, X = 0.9) determined here. Judging. Since the maximum score is smaller than the threshold value X in the document estimation result 15-1 (step ST14 “NO”), the result integrating unit 16 proceeds to the process of step ST16. In step ST16, for each document ID, the weighted addition of the score of the document search result 13-1 and the score of the document estimation result 15-1 is performed to create a final search result 17-1. In FIG. 10, a final search result 74 is obtained by adding (score of document estimation result 15-1) :( score of document search result 13-1) = 1: 1.

On the other hand, when the maximum score of the document estimation result 15-1 exceeds the threshold value X in step ST14 (step ST14 “YES”), the result integration unit 16 discards the document search result 13-1 in the subsequent step ST15. The document estimation result 15-1 is set as a final search result (not shown).
When the search is completed, the document search device displays the title of the document ID and the like on the screen and makes the user select, thereby presenting a desired document position.

  As described above, according to the first embodiment, the document search apparatus receives the search index 5 created from the document 1 prepared in advance and the user input analysis result 11 obtained by analyzing the user input 10, and uses the search index 5. The correspondence between the document search unit 12 that searches the document ID related to the user input analysis result 11 from the document 1 and the assumed question (user utterance) that asks the contents of the document 1 and the document ID that is the answer is defined. The utterance estimation model 9 that has learned the collected utterance data 6, the utterance content estimation unit 14 that estimates the document ID corresponding to the answer to the user input analysis result 11 from the document 1 based on the utterance estimation model 9, and the document search unit 12 Document search result 13 obtained from, and utterance content estimation unit 14 Obtained by integrating the document estimation results 15 was configured with the result integration unit 16 that generates a final search result 17. For this reason, utterance content estimation based on the collected utterance data 6 is performed, which is different from the simple document search function, so that it does not appear in the document 1 input by general users and novice users that could not be realized by the conventional document search function. Search by wording and general terms is possible. Therefore, it is possible to present a more appropriate search result for the user's natural language input than the result of the simple search method.

  Further, according to the first embodiment, the utterance content estimation unit 14 gives a score according to the degree of association with the user input 10 to the estimated document ID, and the result integration unit 16 obtains from the utterance content estimation unit 14. When the score of the obtained document estimation result 15 is larger than a predetermined threshold value X, the document search result 13 obtained from the document search unit 12 is ignored and the final search result 17 is generated. For this reason, in the case of phrases and general terms that do not appear in the document 1 as input by general users and novice users, the simple search method avoids including a large number of inappropriate search result candidates. On the other hand, a more appropriate search result can be presented.

  In the first embodiment, when the maximum score of the document estimation result 15 is larger than the predetermined threshold value X, the document estimation result 15 is used as the final search result 17 as it is. And the score of the document search result 13 may be weighted and added at a predetermined ratio. The score of the document estimation result 15 is calculated from the document directly estimated from the user's utterance, whereas the score of the document search result 13 is calculated from the presence / absence of the keyword in the document. Therefore, there are advantages and disadvantages, respectively, and by adding them by weighting, it is possible to present a score with a good score in both methods.

  Further, according to the first embodiment, the document search apparatus collects the collected utterance data 6 that defines the correspondence between the document 1 prepared in advance and the user utterance that asks the contents of the document 1 and the document ID that is the answer. Using the input analysis unit 2 for analyzing the document, the search index creation unit 4 for creating the search index 5 from the document analysis result 3 output from the input analysis unit 2, and the collected utterance analysis result 7 output from the input analysis unit 2 The utterance estimation model creating unit 8 that learns the correspondence between the user utterance and the document ID and creates the utterance estimation model 9 is provided. For this reason, it is possible to search with words and general terms that do not appear in the document 1 and are input by general users and novice users, which could not be realized by the conventional document search function.

Embodiment 2. FIG.
FIG. 11 is a block diagram showing the configuration of the document search apparatus according to the second embodiment. 11 that are the same as or equivalent to those in FIG. 1 are denoted by the same reference numerals and description thereof is omitted.
Major differences from the first embodiment are the following two points.
(1) Create an utterance estimation model 9 in which the unit of the document ID to which the collected utterance data 6 is assigned is not a fine unit but a larger unit.
(2) The document estimation result 15 is used for the purpose of limiting the search target range by the search index 5.

  In FIG. 11, the search target limiting unit 18 limits the search target of the document search unit 12 to the lower document ID of the document estimation result 15. The document limitation list 19 holds limited document IDs.

  FIG. 12 is a diagram illustrating a hierarchy of document IDs of document 1. In the example of FIG. 12, the collection utterance data 6 is not allocated to the document ID of the lower layer than the second layer (document ID surrounded by the square), but is allocated to the document IDs of the first layer and the second layer. Yes.

Next, the operation of the document search apparatus will be described.
The operation in the creation process is basically the same as in the first embodiment. However, the allocation of the collected utterance data 6 to the document ID is set to the second hierarchy or higher as shown in FIG. Therefore, in FIG. 4, the collected utterance data 6-1 is assigned to the document ID “Id_10”, and all other collected utterance data 6-2 to 6-4 are assigned to “Id_10_1”.

Next, the search process will be described.
FIG. 13 is a flowchart showing the operation from the user input 10 until the final search result 17 is created. FIG. 14 is a diagram for explaining the operation of the search target limiting unit 18. As in the first embodiment, here, the user input 10 is assumed to be a text input, and the user input 10-1 in FIG. 9 is input. In step ST11, the input analysis unit 2 analyzes the user input 10-1 as in FIG. Next, in step ST12, the utterance content estimation unit 14 performs utterance content estimation. The estimation result is the document estimation result 15-2 (document ID, score) in FIG. As described above, since the allocation of the collected utterance data 6 to the document ID is limited to the second hierarchy or higher, there is no document ID in the third hierarchy or lower.

  In subsequent step ST21, the search target limiting unit 18 confirms whether or not there is one or more document IDs for which the score of the document estimation result 15-2 is greater than or equal to a threshold Y (for example, Y = 0.6). In the document estimation result 15-2, since the score of “ID — 10_1” is 0.6 or more (step ST21 “YES”), the process moves to step ST22, and the lower layer of the document ID whose score is greater than or equal to the threshold Y is expanded and expanded. The same score is assigned to each document ID. Further, in the document estimation result 15-2, only “Id — 10_1” is equal to or greater than the threshold Y, so the search target limiting unit 18 selects “Id — 10_1 — 1” to “Id — 10 — 1_7” in the lower layer of “Id — 10_1” as search targets, and the document limitation list 19 Set as -1.

In subsequent step ST23, the document search unit 12 searches the search index 5 using the keyword list 11-2 of FIG. 14, and obtains a document search result 13-1. In step ST24, a result obtained by adding the score of the document restriction list 19-1 to the score of the document search result 13-1 is output as the final search result 17-2.

  On the other hand, if there is no score exceeding the threshold Y in the document estimation result 15-2 in step ST21 (step ST21 “NO”), the search target limiting unit 18 discards the document estimation result 15-2 (step ST25). In step ST26, the document search unit 12 obtains a document search result (not shown) with all document IDs as search targets, and outputs it as a final search result (not shown).

  As described above, according to the second embodiment, the document search apparatus includes a document ID having a score equal to or higher than a predetermined threshold Y among the document estimation results 15 obtained from the utterance content estimation unit 14 and a document ID of a lower layer thereof. The utterance content estimation unit 14 includes a search target limiting unit 18 that extracts the utterance, and the utterance content estimation unit 14 learns the correspondence between the document ID of the higher hierarchy and the collected utterance data 6 as a minimum unit of search of the search index 5. Based on the model 9, the result integration unit 16 obtained from the document search unit 12 the document ID extracted by the search target limiting unit 18 in the document estimation result 15 obtained from the utterance content estimation unit 14. The document search result 13 is integrated. For this reason, if the collected utterance data 6 is assigned to a document ID of a higher hierarchy, it is possible to associate the collected utterance data 6 with a document ID that does not need to take into account fine differences in functions depending on the product model. . Therefore, it becomes easy to associate the document ID with the collected utterance data 6, and it is possible to suppress a decrease in search accuracy due to data sparseness. In addition, since product functions can be defined at a general-purpose level, it can be used as common collected utterance data 6 in product development with many models, and it becomes easy to deal with new products.

In Embodiments 1 and 2 described above, the search index 5 is described using a search index of the statistical search method. However, the search index of the logical search method is used to calculate the total number of occurrences of the search keyword. A probability may be set for each. In this case, N is the case where the total number of search keyword appearances is the maximum, and a score obtained by dividing the total number of search keyword appearances in each document by N is used as a score, or the number of search keyword appearances in all documents in the search results. A method may be considered in which the sum is M , and the score obtained by dividing the total number of search keyword appearances in each document by M is used as a score.

  Furthermore, in the first and second embodiments, the example in which the search index 5 is created and the speech estimation model 9 is created in units of independent words has been described. However, the phoneme n-gram, syllable n-gram, etc. The search index 5 and the utterance estimation model 9 may be created. Alternatively, the search index 5 and the utterance estimation model 9 may be created by combining a frequently-occurring word and a phoneme n-gram, or a highly frequently-occurring word and a syllable n-gram. In this case, the size of the search index 5 and the utterance estimation model 9 can be reduced.

  In Embodiments 1 and 2 described above, utterances that cannot be applied anywhere in document 1 because there is no relevant product function and no appropriate explanation part, such as collected utterance data 6-4 in FIG. When the utterance estimation model 9 is created by giving a special document ID, and the document with the maximum score of the document estimation result 15 with respect to the user input 10 is the special document ID, the result integrating unit 16 The final search result 17 may be created without using the document search result 13. In this case, the document search apparatus may be configured to present a message corresponding to a special document ID.

  Furthermore, in the first and second embodiments, the case where the user input 10 is text input has been described as an example. However, voice recognition may be used as an input unit. In that case, a method of processing the first candidate text of the speech recognition result as the user input 10 and a method of processing up to the Nth candidate as the user input 10 are conceivable. If the speech recognition result is generated in units of morphemes, the processing in the input analysis unit 2 may be omitted and handled as the user input analysis result 11 as it is.

  In the first and second embodiments, examples of Japanese input have been described. However, the language is not limited, and the input analysis unit 2 is replaced for each language in English, German, Chinese, and the like. It is possible to obtain the same effect.

Embodiment 3 FIG.
In the following, an example of English input will be described.
Since the document search apparatus according to the third embodiment has the same configuration as that of the document search apparatus shown in FIG. 1, the following description will be given with reference to FIG.

  FIG. 15 is an English example of the document 1 input to the document search apparatus according to the third embodiment. The document 1 has a hierarchical structure such as a chapter, a section, and an item, and has a document ID indicating a search result position for each hierarchy. In the example of FIG. 15, the document 1-11 having the document ID “Id — 10_1” also includes text included in the lower data structure. For example, the document 1-12 of “Id_10_1_1” is included in the document 1-11 of “Id_10_1”.

FIG. 16 is an example of a keyword list for the document analysis result 3 and the search index 5. “Id — 10 — 1 — 1” is an example of the document analysis result, and shows the result of performing input analysis by morphological analysis on the document 1-12 of “Id — 10 — 1 — 1” in FIG. In this document analysis result 3-11, only information obtained by dividing the morphological analysis result by “/” is presented, but actually information such as part-of-speech information is also generated.
The search index data 3-12 shows an example of data used by the search index creation unit 4 based on the document analysis result 3-11 of “Id — 10_1 — 1”. Here, document IDs and independent word morphemes excluding prepositions, articles, be verbs, and pronouns are extracted.

  FIG. 17 is an example of the collected utterance data 6. The collected utterance data 6-11 is an example of a question corresponding to the document “Id_10”, the collected utterance data 6-12 is an example of a question corresponding to the document of “Id_10_1”, and the collected utterance data 6-13 is a document of “Id_10_1_1”. It is an example of a question corresponding to. The collected utterance data 6-14 is a question intended to know a specific method of changing the map type. However, since it is a map type that cannot be realized with the product assumed here, it is the same as “Id_10_1_1”. It is an example of the collection utterance data which cannot select document ID of a hierarchy.

  FIG. 18 is an example of the keyword list for the collected utterance analysis result 7 and the utterance estimation model 9. The collected utterance analysis result 7-11 of "Id_10_1_1" is the collected utterance analysis result example of the collected utterance data 6-13 of "Id_10_1_1" in FIG. The example of data which the speech estimation model preparation part 8 based on the result 7-11 uses is shown. Here, document IDs and independent word morphemes excluding prepositions, articles, and be verbs are extracted.

Next, the operation of the document search apparatus will be described.
The operation (creation process, search process) of the document search apparatus according to the third embodiment is basically the same as that in FIGS. 6 to 8 of the first embodiment. Therefore, only different parts will be described here. First, the creation process will be described.

First, a method for creating the search index 5 in the creation process will be described. Here, weighting by tf · idf disclosed in the prior art is assumed.
As shown in FIG. 15, it is assumed that the document 1 is a pair in which the document ID and the text are associated with each other. For example, in the document 1-12, the text “Heading up. Display the map when to face faces the direction you are traveling” is associated with the document ID “Id — 10 — 1 — 1”. In step ST1 of FIG. 6, the input analysis unit 2 sequentially reads the document 1 having this structure and divides it into morpheme strings by morphological analysis which is a known technique. The result of the morphological analysis of the document 1-12 is a document analysis result 3-11 in FIG. This document analysis result 3-11 shows only morpheme breaks, but in reality, it is assumed that part-of-speech information, the original form of the usage word, and the like are included.

  When the document analysis result 3 is generated for all document IDs, the search index creation unit 4 extracts morphemes (keywords) necessary for creating the search index 5 from all the document analysis results 3 in the subsequent step ST2. Then, a pair of (document ID, keyword list) is created, and a search index 5 weighted by tf · idf based on all pairs is created. A pair of (document ID, keyword list) extracted from the document analysis result 3-11 in FIG. 16 is represented by the same search index data 3-12 in FIG.

  The specific search index creation procedure is the same as that in the first embodiment, and a description thereof will be omitted.

Next, a process for creating the utterance estimation model 9 will be described.
The collected utterance data 6 is data in which utterances collected from the user in advance are assigned to the document ID as the answer, as shown in the collected utterance data 6-11 to 6-14 in FIG. Since the method of creating the collected utterance data 6 is the same as that in the first embodiment, description thereof is omitted.

  In step ST3 shown in FIG. 7, the input analysis unit 2 performs morphological analysis on the collected utterance data 6 in the same manner as when the document 1 is input in step ST1 described above. For example, the result of morphological analysis of the collected utterance data 6-13 in FIG. 17 is the collected utterance analysis result 7-11 in FIG. In subsequent step ST4, the utterance estimation model creating unit 8 extracts a list of document IDs and keywords as utterance estimation model data 7-12 in the same manner as in step ST2 described above, and the ME method as in the first embodiment. To learn the utterance estimation model 9. Keywords are extracted from all the collected utterance analysis results 7 and learned by the ME method to create an utterance estimation model 9. Specifically, the same utterance estimation model data 7-12 of FIG. 18 is extracted from the collected utterance analysis result 7-11 of FIG. 18, and the above learning is performed based on the utterance estimation model data 7-12. Done.

Next, the search process will be described.
19 and 20 show a transition example in the search process of the user input 10-11 which is an example of the user input 10. FIG. Here, it is assumed that the user input 10 is a text input and the user input 10-11 in FIG. 19 is input. The input analysis unit 2 first receives the user input 10-11 in step ST11 shown in FIG. 8 and generates a user input analysis result 11-11 by performing morphological analysis. From the user input analysis result 11-11, a preposition, an article, Excludes be verbs and pronouns, extracts independent words, and creates a keyword list 11-12. In the subsequent step ST12, the utterance content estimation unit 14 uses the keyword list 11-12 as an input to obtain the document estimation result 15-11 of FIG. As shown in FIG. 20, the document estimation results 15-11 are arranged in the order of scores.

  When the document estimation result 15-11 is obtained, the document search unit 12 obtains the document search result 13-11 of FIG. 20 from the search index 5 by using the keyword list 11-12 as an input in the next step ST13. As shown in FIG. 20, the document search results 13-11 are also arranged in the order of score.

  In subsequent step ST14, the result integration unit 16 determines whether or not the maximum score of the document estimation result 15-11 is equal to or greater than a threshold value X (for example, X = 0.9) determined here. Since the maximum score is smaller than the threshold value X in the document estimation result 15-11 (step ST14 “NO”), the result integration unit 16 proceeds to the process of step ST16. In step ST16, for each document ID, the weighted addition of the score of the document search result 13-11 and the score of the document estimation result 15-11 is performed to create the final search result 17-11. In FIG. 20, the result obtained by adding (score of document estimation result 15-11) :( score of document search result 13-11) = 1: 1 is the final search result 17-11.

On the other hand, when the maximum score of the document estimation result 15-11 exceeds the threshold value X in step ST14 (step ST14 “YES”), the result integration unit 16 discards the document search result 13-11 in the next step ST15. Thus, the document estimation result 15-11 is set as a final search result (not shown).
When the search is completed, the document search device displays the title of the document ID and the like on the screen and makes the user select, thereby presenting a desired document position.

As described above, according to the third embodiment, the document search apparatus can perform the same processing as that of the first embodiment not only on the Japanese language but also on the English document 1. The same effect as in the first embodiment can be obtained.
Although not described, the configuration of the third embodiment may be applied to the second embodiment.

Embodiment 4 FIG.
In the following, an example of Chinese input will be described.
The document search apparatus according to the fourth embodiment has the same configuration as that of the document search apparatus shown in FIG. 1, and therefore will be described below with reference to FIG.

  FIG. 21 is a Chinese example of the document 1 input to the document search apparatus according to the fourth embodiment. The document 1 has a hierarchical structure such as a chapter, a section, and an item, and has a document ID indicating a search result position for each hierarchy. In the example of FIG. 21, the document 1-21 having the document ID “Id — 10_1” also includes text included in the lower data structure. For example, the document 1-22 of “Id_10_1_1” is included in the document 1-21 of “Id_10_1”.

FIG. 22 is an example of the keyword list for the document analysis result 3 and the search index 5. “Id — 10 — 1 — 1” is an example of the document analysis result, and shows the result of performing input analysis by morphological analysis on the document 1-22 of “Id — 10 — 1 — 1” in FIG. In this document analysis result 3-21, only information obtained by dividing the morphological analysis result by “/” is presented, but actually information such as part of speech information is also generated.
The search index data 3-22 is an example of data used by the search index creation unit 4 based on the document analysis result 3-22 of “Id — 10_1 — 1”. Here, the document ID and independent word morphemes excluding pronouns, particles, and interpositions are extracted.

  FIG. 23 is an example of the collected utterance data 6. The collected utterance data 6-21 is an example of a question corresponding to the document “Id_10”, the collected utterance data 6-22 is an example of a question corresponding to the document of “Id_10_1”, and the collected utterance data 6-23 is a document of “Id_10_1_1”. It is an example of a question corresponding to. The collected utterance data 6-24 is a question intended to know a specific method of changing the map type. However, since it is a map type that cannot be realized with the product assumed here, it is the same as “Id — 10_1 — 1”. It is an example of the collection utterance data which cannot select document ID of a hierarchy.

  FIG. 24 is an example of a keyword list for the collected utterance analysis result 7 and the utterance estimation model 9. The collected utterance analysis result 7-21 of "Id_10_1_1" is the collected utterance analysis result example of the collected utterance data 6-23 of "Id_10_1_1" in FIG. 23, and the utterance estimation model data 7-22 is the collected utterance analysis of "Id_10_1_1". The example of data which the speech estimation model preparation part 8 based on the result 7-21 uses is shown. Here, the document ID and independent word morphemes excluding pronouns, particles, and interpositions are extracted.

Next, the operation of the document search apparatus will be described.
The operations (creation processing and search processing) of the document search apparatus according to the fourth embodiment are basically the same as those in FIGS. 6 to 8 in the first embodiment. Therefore, only different parts will be described here. First, the creation process will be described.

First, a method for creating the search index 5 in the creation process will be described. Here, weighting by tf · idf disclosed in the prior art is assumed.
As shown in FIG. 21, it is assumed that the document 1 is a pair in which the document ID and the text are associated with each other.

  In step ST1 of FIG. 6, the input analysis unit 2 sequentially reads the document 1 having this structure and divides it into morpheme strings by morphological analysis which is a known technique. The result of the morphological analysis of the document 1-22 is a document analysis result 3-21 in FIG. This document analysis result 3-21 shows only morpheme breaks, but it is assumed that part of speech information is actually included.

  When the document analysis result 3 is generated for all document IDs, the search index creation unit 4 extracts morphemes (keywords) necessary for creating the search index 5 from all the document analysis results 3 in the subsequent step ST2. Then, a pair of (document ID, keyword list) is created, and a search index 5 weighted by tf · idf based on all pairs is created. A pair of (document ID, keyword list) extracted from the document analysis result 3-21 in FIG. 22 is represented by the same search index data 3-22 in FIG.

  The specific search index creation procedure is the same as that in the first embodiment, and a description thereof will be omitted.

Next, a process for creating the utterance estimation model 9 will be described.
The collected utterance data 6 is data in which the utterances collected from the user in advance are assigned to the document ID as the answer, as shown in the collected utterance data 6-21 to 6-24 in FIG. Since the method of creating the collected utterance data 6 is the same as that in the first embodiment, description thereof is omitted.

  In step ST3 shown in FIG. 7, the input analysis unit 2 performs morphological analysis on the collected utterance data 6 in the same manner as when the document 1 is input in step ST1 described above. For example, the result of morphological analysis of the collected utterance data 6-23 in FIG. 23 is the collected utterance analysis result 7-21 in FIG. In subsequent step ST4, the utterance estimation model creating unit 8 extracts a list of document IDs and keywords as utterance estimation model data 7-22 in the same manner as in step ST2 described above, and the ME method as in the first embodiment. To learn the utterance estimation model 9. Keywords are extracted from all the collected utterance analysis results 7 and learned by the ME method to create an utterance estimation model 9. Specifically, the same utterance estimation model data 7-22 of FIG. 24 is extracted from the collected utterance analysis result 7-21 of FIG. 24, and the above learning is performed based on the utterance estimation model data 7-22. Done.

Next, the search process will be described.
FIG. 25 and FIG. 26 show a transition example in the search process of the user input 10-21 which is an example of the user input 10. Here, it is assumed that the user input 10 is a text input and the user input 10-21 in FIG. 25 is input. The input analysis unit 2 first receives the user input 10-21 in step ST11 shown in FIG. 8, generates a user input analysis result 11-21 by performing morphological analysis, and generates a pronoun, particle, Independent words are extracted by excluding interverbs, and a keyword list 11-22 is created. In the subsequent step ST12, the utterance content estimation unit 14 uses the keyword list 11-22 as an input to obtain the document estimation result 15-21 of FIG. As shown in FIG. 26, the document estimation results 15-21 are arranged in the order of scores.

  When the document estimation result 15-21 is obtained, in the subsequent step ST13, the document search unit 12 obtains the document search result 13-21 of FIG. 26 from the search index 5 by using the keyword list 11-22 as an input. As shown in FIG. 26, the document search results 13-21 are also arranged in the order of score.

  In subsequent step ST14, the result integration unit 16 determines whether or not the maximum score of the document estimation result 15-21 is equal to or greater than a threshold value X (for example, X = 0.9) determined here. Since the maximum score is smaller than the threshold value X in the document estimation result 15-21 (step ST14 “NO”), the result integration unit 16 proceeds to the process of step ST16. In step ST16, for each document ID, the weighted addition of the score of the document search result 13-21 and the score of the document estimation result 15-21 is performed to create the final search result 17-21. In FIG. 26, the final search result 17-21 is obtained by adding (score of document estimation result 15-21) :( score of document search result 13-21) = 1: 1.

On the other hand, when the maximum score of the document estimation result 15-21 exceeds the threshold value X in step ST14 (step ST14 “YES”), the result integration unit 16 discards the document search result 13-21 in step ST15. Thus, the document estimation result 15-21 is set as a final search result (not shown).
When the search is completed, the document search device displays the title of the document ID and the like on the screen and makes the user select, thereby presenting a desired document position.

As described above, according to the fourth embodiment, the document search apparatus can perform the same processing as that of the first embodiment on not only Japanese but also the Chinese document 1. Also, the same effect as in the first embodiment can be obtained.
Although not described, the configuration of the fourth embodiment may be applied to the second embodiment.

  In addition to the above, within the scope of the invention, the invention of the present application can be freely combined with each embodiment, modified any component of each embodiment, or omitted any component in each embodiment. Is possible.

  As described above, the document search apparatus according to the present invention uses the utterance estimation model in which the correspondence between the question assuming the user's way of listening and the document item as the answer is learned, Since the search result in the document with respect to the input in natural language is presented, it is suitable for use in, for example, an information device that searches and displays electronic instruction manuals such as home appliances and in-vehicle devices.

  1 document, 2 input analysis unit, 3 document analysis result, 4 search index creation unit, 5 search index, 6 collected utterance data, 7 collected utterance analysis result, 8 utterance estimation model creation unit, 9 utterance estimation model, 10 user input, 11 user input analysis result, 12 document search unit, 13 document search result, 14 utterance content estimation unit, 15 document estimation result, 16 result integration unit, 17 final search result, 18 search target limiting unit, 19 document limited list.

Claims (6)

A search index created from documents prepared in advance,
A document search device comprising a document search unit that receives an input from a user and searches for an item related to the user input from within the document using the search index,
An utterance estimation model that learns the correspondence between an assumed question asking the content of the document and an item in the document that is an answer to the assumed question;
An utterance content estimation unit that estimates an item corresponding to the answer of the user input from within the document based on the utterance estimation model;
A document search apparatus comprising: a result integration unit that integrates a document search result obtained from the document search unit and a document estimation result obtained from the utterance content estimation unit to generate a final search result. The utterance content estimation unit gives a score according to the degree of association with the user input to the estimated item in the document,
The result integration unit generates a final search result by ignoring the document search result obtained from the document search unit when the score of the document estimation result obtained from the utterance content estimation unit is larger than a predetermined value. The document search apparatus according to claim 1, wherein: The document search unit gives a score corresponding to the degree of association with the user input to the items in the searched document,
The utterance content estimation unit gives a score according to the degree of association with the user input to the estimated item in the document,
The result integration unit adds and integrates a score of a document search result obtained from the document search unit and a score of a document estimation result obtained from the utterance content estimation unit at a certain ratio. Item 2. The document search device according to Item 1. Among the document estimation results obtained from the utterance content estimation unit, a search target limiting unit for extracting items satisfying a predetermined condition,
The utterance content estimation unit estimates based on an utterance estimation model in which a correspondence relationship between an item of a unit larger than the minimum unit of search of the search index and the assumed question is learned,
The result integration unit integrates the items extracted by the search target limiting unit from the document estimation results obtained from the utterance content estimation unit with the document search results obtained from the document search unit. The document search apparatus according to claim 1. An input analysis unit that analyzes collected utterance data that defines a correspondence relationship between a prepared document and an assumed question asking the content of the document and an item in the document that is an answer to the question;
A search index creation unit for creating the search index from the analysis result of the document output from the input analysis unit;
Using the analysis result of the collected utterance data output from the input analysis unit, learning a correspondence relationship between the assumed question and the item in the document, and an utterance estimation model creation unit that creates the utterance estimation model; The document search apparatus according to claim 1, further comprising: A document search method using a document search device,
A user input step in which the input analysis unit receives input from the user;
Document retrieval unit, using the search index created from previously prepared document, and document retrieval step of retrieving the items that are relevant to the user input from the document,
Based on the utterance estimation model in which the utterance content estimation unit has learned the correspondence between the assumed question asking the content of the document and the item in the document that is the answer to the assumed question, the answer of the user input from within the document An utterance content estimation step for estimating an item corresponding to
A document search method comprising: a result integration step in which a result integration unit generates a final search result by integrating a document search result obtained from the document search step and a document estimation result obtained from the utterance content estimation step.

Images